Banana Leaf Disease Dataset of Bangladesh
Description
This dataset provides a real-world collection of 2,638 raw, unprocessed images (.jpg) of banana leaves, gathered from various regions across Bangladesh. The data is categorized into three classes: Healthy (638 images), Panama Disease (497 images), and Sigatoka Disease (1,503 images). To facilitate immediate machine learning use, it is pre-partitioned into Training (1,603 images), Validation (517 images), and Testing (518 images) subsets. Unlike highly controlled plant datasets, this collection preserves the raw pixel-wise variations found in authentic agricultural environments, providing a rigorous benchmark for computer vision systems. Value of the Data: Real-World Variability: Many existing datasets feature images taken in controlled lab settings. This dataset captures genuine pixel-wise differences, including drastic variations in natural lighting, diverse camera angles, shadowing, and complex natural backgrounds. Unprocessed and Raw: Images are provided in their native state without computational preprocessing or resizing. This allows researchers to apply custom data augmentation or normalization pipelines without inherited artifacts. Model Robustness: Reflecting the messy reality of the field, it is an excellent benchmark for testing the true generalization capabilities of deep learning models and Explainable AI (XAI) techniques. Data Description: The main directory "banana_ld_dataset.zip" contains "train", "val", and "test" folders. Each subset is subdivided into the three biological classes. The exact distribution is as follows: Training Set (1,603 images): Healthy: 399 Panama: 299 Sigatoka: 905 Validation Set (517 images): Healthy: 119 Panama: 99 Sigatoka: 299 Testing Set (518 images): Healthy: 120 Panama: 99 Sigatoka: 299 Experimental Design, Materials, and Methods Images were collected directly from active banana plantations across diverse regions of Bangladesh to capture a wide representation of the local agricultural landscape. Photographs were taken using standard digital cameras under highly variable natural lighting, capturing infected and healthy leaves from multiple angles and distances. To maintain the integrity of real-world conditions, the images were visually diagnosed and sorted by phytopathological class but intentionally left in their native .jpg format. No cropping, color correction, background removal, or other preprocessing techniques were applied prior to publication, ensuring the data remains entirely raw.
Files
Steps to reproduce
Since this dataset intentionally provides raw, unprocessed images to preserve real-world variability, researchers utilizing this data for machine learning or deep learning applications are advised to implement the following preprocessing pipeline: 1. Image Resizing and Standardization The images are provided in their original resolutions. Depending on the target architecture (e.g., ResNet, MobileNetV2, or Vision Transformers), all images should be resized to a uniform dimension. Standard resolutions such as 224x224 or 256x256 pixels are highly recommended to balance computational efficiency with the retention of fine pathological details. 2. Pixel Normalization To ensure stable model convergence, the raw pixel values (which range from 0 to 255) must be normalized. Researchers should scale the pixel values to a range of [0, 1] by dividing by 255.0. Alternatively, applying Z-score standardization using ImageNet dataset mean and standard deviation weights is recommended if utilizing pre-trained transfer learning models. 3. Data Augmentation Because the dataset captures genuine agricultural conditions with variable lighting and angles, robust data augmentation is encouraged during the training phase to prevent overfitting. Recommended augmentations include: Random horizontal and vertical flipping. Random rotations (e.g., ±20 degrees). Color jittering (slight adjustments to brightness, contrast, and saturation) to simulate different times of day and weather conditions. 4. Addressing Class Imbalance Users should note the natural class distribution within the dataset (Healthy: 638, Panama: 497, Sigatoka: 1,503). To prevent models from biasing toward the majority class (Sigatoka), researchers should implement imbalance-handling techniques. Suggestions include using weighted loss functions (e.g., Class-Balanced Cross Entropy), focal loss, or applying oversampling/undersampling techniques within the data loader. 5. Framework Integration The dataset’s directory structure (train, val, test with nested class folders) is fully compatible with standard deep learning data loaders. Users can directly mount the dataset using functions like tf.keras.utils.image_dataset_from_directory in TensorFlow/Keras or torchvision.datasets.ImageFolder in PyTorch without needing to write custom parsing scripts.
Institutions
- American International University-BangladeshDhaka Division, Dhaka